Evaluation of <scp>BugBox</scp> , a software platform for <scp>AI</scp> ‐assisted bioinventories of arthropods

Artificial intelligence (AI) technology has the potential to revolutionize entomology and biodiversity research, allowing entomologists to address biodiversity questions on a larger scale than ever before. A new software program, called BugBox, has been developed to facilitate large-scale arthropod bioinventories. BugBox uses an AI algorithm to rapidly classify arthropods from specimen photographs and calculates per-sample diversity indices from its classifications. We evaluated the performance of the AI algorithm over three consecutive training cycles by comparing the AI's classifications to identifications conducted by a human expert. We also used both AI and human data to separately test the hypothesis that regenerative agricultural practices increase arthropod biodiversity in a bioinventory from North American rangelands. BugBox demonstrated substantial improvement in all test metrics over the three cycles as it was allowed to incorporate the human expert's corrections into each new model version (e.g. f1 score improved from 0.523 to 0.722 over the four consecutive model versions). AI classifications were strongly correlated with human identifications, and the AI drew the same conclusion as the human data when comparing diversity indices (Hill numbers): both found evidence that regenerative practices increased arthropod diversity. These results demonstrate that, while the AI was less accurate than the human, it was still able to provide useful surrogate data at scale very rapidly. It can also improve over time under the guidance of human expertise. This technology has profound implications for the scalability of entomological science.